【作者】 孟飚；

【导师】 张雪英；

【作者基本信息】 太原理工大学 ， 信号与信息处理， 2003， 硕士

【摘要】 本文先介绍了语音编码的发展情况，之后详细地论述了CS-ACELP算法，即8kbit/s的共轭结构代数码激励线性预测编码的完整结构和算法，对包括预处理、线性预测分析和量化(加窗和自相关计算、Levinson-Durbin算法、LP到LSP的转换、LSP系数的量化、LSP系数的内插、LSP到LP的转换)、感觉加权、开环基音分析、冲激响应计算、目标信号的计算、自适应码书的搜索(自适应码书矢量的产生、延迟码字的计算)、固定码书的构造和搜索、增益的量化(增益预测、用于增益量化的码书搜索、用于增益量化的码字计算)、存储器更新，语音合成后的处理等各模块的功能和理论基础作了细致分析。并对涉及到的语音处理的关键技术，如线性预测、LPC与LSP的转换、矢量量化、基音分析等技术作了深入研究。用标准C语言仿真实现了该算法，计算了MOS分值，女声：4.180497，男声：4.199782，并在相同的测试语句中加入噪声进行测试，含噪语句通过该编解码器，输出的合成语音用主、客观评价标准评价，与原始不含噪语音效果差别不大，平均MOS分值为：女声4.1375，男声4.1668，说明该算法是优秀的编解码算法。 此外，特别就CS-ACELP算法中的LSP量化方面作了深入的研究，尝试了几种不同的量化方法：(1)改变分裂式矢量量化的维数组合，原算法中第二级残差量化时用了两段式分裂量化法，将10维矢量分裂为两个5维矢量。本研究中，通过实验发现3维-7维的分法效果最好。(2)进行了码书优化。对LSP参数量化中的第一级码书的128个码字的使用频率进行了统计试验，选用了128个码字中使用频率高的112个码字作为新码书，语音质量基本不变但降低了码书搜索的复杂度。更多还原

【Abstract】 This paper describes the 8kbit/s speech coding algorithm which has been standardized by ITU-T in 1996. The algorithm is based on a Conjugate-Structure Algebraic Code Excited Linear Prediction (CS-ACELP) coding technique and uses 1 Oms (80 samples at an 8 kHz sample rate) speech frames. This coder will be used for the Future Public Land Mobile Telecommunication System and will be suitable for Personal Communication Service. The coder delivers toll-quality speech (equivalent to 32kbit/s ADPCM) for most operating conditions. The coder operates on speech frames of 10ms, computes the long-term predictor coefficients, and operates in an analysis-by-synthesis loop to find the excitation vector that minimizes the perceptually weighted error signal.In this paper, the coder structure is described, the algorithm about CS-ACELP is discussed, and its central aspects are analyzed in detail. To achieve high-quality speech and real-time implementation, CS-ACELP has been revised by novel schemes. Efficient pitch and codebook search strategies, along with efficient quantization procedures, have been developed to achieve toll quality encoded speech. LSP parameters are quantized by multi-stage VQ with fourth-order interframe MA prediction. This scheme has little spectrum distortion, even if the two types of speech have many variations of LSP parameters. Moreover, computational complexity for implementation is reduced in adaptive and fixed-shape codebooks without degrading the quality. Multi-stage selection is adopted in the adaptive codebook; this selection uses atruncated impulse response. Improved pre-selection is proposed in the fixed-shape codebook. Subjective testing indicates that the quality of CS-ACELP is equivalent to that of the 32kbit/s Adaptive Differential Pulse Code Modulation (ADPCM) under error-free conditions and it outperforms G.726 under error condition.In this paper, Standard C is adopted in realization of the algorithm, presents program strategies and steps of algorithm of each module. The coder and decoder is tested by utterances with noise. The results are satisfying. Moreover, the paper studies the quantization of LSP and tries some other means to quantize the LSP parameter. We adopt the means as follows: first, we change the splitting dimension of the second grade codebook which was splirted into double five dimension. By testing, we find a better splitting way. Second, we optimize the codebook and choice a part of the codeword which is used most efficiently. The result is not degraded too much while the complexity is reduced. At the end of the paper the development prospect of CS-ACELP and speech coding are described.更多还原